Efficient lens design enabled by a multilayer perceptron-based machine learning scheme

Abstract

Machine learning is a major branch of artificial intelligence that has been widely implemented in optical applications. Here we establish and validate two fully connected feed forward artificial neural networks (ANNs) based on the multilayer perceptron incorporating double hidden layers. The constructed ANNs act as regressors to efficiently predict the divergence and deflection angles of beams emerging from the beam-converging and -deflecting lenses, respectively. The target lens specifications can then be inversely queried by the desired beam divergence and deflection angles contained in the forecasted datasets. With the aid of meticulous hyperparameter tuning, the optimized ANNs of the beam-converging and -deflecting lenses yield high coefficient of determination (R2) scores of 9.9964e–1 and 9.9933e–1, and low mean squared error (MSE) losses of 1.0e–5 and 2.2e–5, respectively. Compared with the conventional optical design, the proposed scheme has been confirmed to substantially alleviate the complexity of lens design, provide rich lens specification solutions for different beam divergence and deflection angles, and drastically reduce the computation time by over four orders of magnitude.